daily-held: #1689 Phase-D disposition — descoped CIs + 44h w
kind: infra#daily-held#needs-human
track:Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Held decision (needs Thomas)
Filed by /daily 2026-07-27 problem sweep as a route-3 judgment call. Carve-out item: scientific-meaning change + spends money/compute
#1689's Phase-D ladder bootstrap was descoped mid-run without a plan amendment:
headline CIs now rest on 10 bootstrap / 2 null draws (pre-registered: 1000/40;
R15-marker-documented: 200/40) and L19-only vs 4 planned layers. The R16
vectorized port (torch fp64, 16 workers, GPUs now saturated) still projects
44h more wall on the 4xH100 pod ($10-12/hr class), documented continue_as_is
in compute-deviation v4. Decisions only you can make:
- Accept the descoped draws for the clean-result (with a planned-vs-actual caveat), or order a re-bootstrap at B>=1000 on the winning pairs (now cheap after the ~10x vectorization) before promotion.
- Keep the 4xH100 pod for the remaining ~44h vs persist + downsize.
- The instruct-leg point estimates are FULL quality regardless of draws (computed once pre-draw-loop) — a prior in-chat recommendation to terminate that leg was wrong and withdrawn. Evidence: #1689 epm:compute-deviation v3/v4, epm:progress v64 audit, gpu-idle-escalation x3; miners A P1, B P2, I P1-P4, J P1. Suggested action: answer 1-2 on the task; the owning session executes.